Home › Articles

Turning-Point Prediction, What Is Special About the Nasdaq-100/QQQ, and the Realities of Retail Implementation

2026-10-26 · Research note · Series: VOO/QQQ market-timing evidence (part 5 of 5) · Not investment advice; figures are as of the sources and dates cited in the text · 中文
Contents6 sections
  1. Q1 Can academic research predict bull/bear turning points (tops/bottoms) in real time?
  2. Q2 Evidence specific to the Nasdaq-100/QQQ and the risk of extrapolating from the SP 500
  3. Q3 Retail implementation realities: trading costs, taxes, and investors' actual timing record
  4. Q4 Free data sources and publication lags
  5. Further reading
  6. Frequently asked questions

This note covers two areas the general equity premium prediction literature does not address: (1) predicting bull/bear "turning points" (tops and bottoms), rather than predicting average returns; (2) evidence specific to the Nasdaq-100/QQQ, plus the costs, taxes and data sources a retail investor actually faces when implementing. The baseline prior is "timing does not work" (Goyal & Welch 2008; Goyal, Welch & Zafirov 2024 RFS).

Site note: This is one of five articles in the series "Ranking the Empirical Evidence on VOO/QQQ Timing Methods" (this article: turning points and Nasdaq implementation). It is the full text of the site owner's research note, published as is: the content, numbers, tables, citations and "not found / could not be verified" passages appear exactly as in the original, with nothing cut or rewritten; only the table-of-contents block at the top of the original was removed. The source file is turning_points_nasdaq_implementation.md in the series folder (VOO/QQQ timing methods: empirical ranking) under the project's research_notes/ directory. The original of this article does not state a research date; the other notes in the series give a research date of 2026-10-26. The default prior throughout the series is "timing does not work" (Goyal & Welch 2008; Goyal, Welch & Zafirov 2024). Verification levels follow the original definitions: [Verified: original] = read directly in the paper's original text (with page/table number); [Verified: abstract] = read the journal or official abstract/bibliographic record, with numbers taken from the abstract; [Secondary] = numbers come from secondary write-ups (blogs, news) and serve only as leads; [Not verified page by page] = from the researcher's existing knowledge, original not obtained, specific numbers must be checked against the original. This article is a research record, not investment advice.

Verification level tags:

  • [Verified: original]: read directly in the paper's original text this time (with page/table number)
  • [Verified: abstract]: read the journal or official abstract/bibliographic record this time; numbers come from the abstract
  • [Secondary]: numbers come from secondary write-ups (blogs, news) and serve only as leads
  • [Not verified page by page]: from the researcher's existing knowledge; the original was not obtained this time; the bibliographic record and DOI were confirmed as far as possible, but specific numbers must be checked against the original

Q1 Can academic research predict bull/bear turning points (tops/bottoms) in real time?

Takeaway

There is no credible evidence that any method can "catch" the market's top or bottom in real time. What the literature can do: (a) date bull/bear cycles after the fact; (b) predict the probability that "next month will be in a bear state", where some models have out-of-sample (OOS) evidence of beating a constant probability, and a few small timing exercises beat buy-and-hold, but these are mostly single samples, point estimates, before tax, and not independently replicated afterwards; (c) the finding that "crash probability rises" after an industry run-up is robust, but the timing of the crash cannot be judged in advance, and the original authors admit their trading-rule threshold has look-ahead bias. Bubble detection (GSADF) "detects that an explosive run-up is already under way"; it does not predict the top.

Cited Findings

Baseline prior (predicting average returns)

  • Goyal & Welch (2008) find that none of the variables in the pre-2005 literature could consistently beat the historical average equity premium out of sample: "not a single one would have helped a real-world investor". RFS 21(4):1455–1508, DOI 10.1093/rfs/hhm014 [Verified: abstract] — Causal Claims summary page
  • Goyal, Welch & Zafirov (2024) retest 29 new variables from 26 papers published after 2008 plus the original 17 variables (data through the end of 2021): more than one-third of the new variables are no longer significant even in sample; about half of those still significant in sample perform poorly out of sample; only a few pass both. RFS 37(11):3490–3557, DOI 10.1093/rfs/hhae044 [Verified: abstract] — SFI page; Ivo Welch paper card

Dating bull/bear cycles (only ex-post definitions, not predictions)

  • Pagan & Sossounov (2003) "A simple framework for analysing bull and bear markets" adapts the Bry-Boschan business-cycle dating algorithm to monthly stock prices (commonly cited parameters: an 8-month window for identifying peaks and troughs, each phase at least 4 months, a complete cycle at least 16 months, and a fall/rise of more than 20% can override the minimum-duration rule). JAE 18(1):23–46, DOI 10.1002/jae.664 [Not verified page by page; parameters are commonly cited values and must be checked against Section 2 of the original] — Wiley DOI
    • Key point: rules like this can only confirm a top at least several months "after" it, so they are by nature ex-post dating and cannot call the top in real time.
  • Lunde & Timmermann (2004) "Duration Dependence in Stock Prices: An Analysis of Bull and Bear Markets": uses hazard models to describe how the probability of a bull or bear market ending relates to its elapsed duration and to interest rates; rejects the random walk; GARCH does better but is still inconsistent with very long bull markets; rising interest rates lower the hazard of a bear market ending (bear markets are more likely to persist). JBES 22(3):253–273 [Verified: abstract] — IDEAS/RePEc working paper page; CEPR DP4104
    • This is a descriptive/in-sample analysis of bull and bear durations, not an OOS trading test.
  • Maheu & McCurdy (2000) "Identifying Bull and Bear Markets in Stock Returns", Markov switching with duration dependence, JBES 18(1):100–112, DOI 10.1080/07350015.2000.10524851 [Not verified page by page; mainly in-sample identification] — DOI
  • Maheu, McCurdy & Song (2012) "Components of Bull and Bear Markets: Bull Corrections and Bear Rallies", four-state Markov switching, JBES 30(3):391–403, DOI 10.1080/07350015.2012.680412 [Not verified page by page; the abstract as remembered includes an OOS density forecast evaluation, but there are no verified hit-rate/trading-performance numbers] — DOI

Predicting bull/bear "states" (studies with OOS evidence)

  • Chen (2009) "Predicting the bear stock market: Macroeconomic variables as leading indicators" (note: the original title is not "Predicting recessions"): monthly S&P 500 data, testing the term spread, inflation, money, output, unemployment, the federal funds rate, government debt and exchange rates; the yield-curve spread and inflation are the most useful bear-market predictors, both in sample and out of sample; predicting bear markets with macro variables is easier than predicting returns. JBF 33(2):211–223, DOI 10.1016/j.jbankfin.2008.07.013 [Verified: abstract] — NTU Scholars
    • The sample period 1957/02–2007/12 and the sign of each variable (spread, inflation and unemployment positively related to bear probability, industrial production growth negatively related) come from a secondary write-up [Secondary] — CXO Advisory
    • Specific OOS QPS/hit-rate numbers: insufficient evidence / could not be verified.
  • Nyberg (2013) "Predicting bear and bull stock markets with dynamic binary time series models": US monthly data; bull and bear states are predictable both in sample and out of sample; adding dynamic structure (lagged state/lagged probability) improves predictive power; a small timing exercise allocating between stocks and bonds based on the probability forecasts returns more than buy-and-hold. JBF 37(9):3351–3363, DOI 10.1016/j.jbankfin.2013.05.008 [Verified: abstract; sample period, QPS, size of excess return, and whether costs are deducted: could not be verified] — IDEAS/RePEc; Helsinki research portal
    • Note: in a dynamic probit the "lagged state" itself must be determined by an ex-post dating rule, and dating requires a confirmation period of several months; if handled carelessly this introduces look-ahead problems. Whether state variables available in real time were used could not be verified this time.
  • Kole & van Dijk (2017) "How to Identify and Forecast Bull and Bear Markets?": S&P 500, comparing rule-based methods (Pagan-Sossounov type) with Markov switching, evaluated by the mean-variance utility of a risk-averse investor; rule-based methods are better for (in-sample) identification, Markov switching is better for (out-of-sample) forecasting; the reason is that in sample only the mean return matters, whereas for forecasting the variance (volatility) is the key input. JAE 32(1):120–139, DOI 10.1002/jae.2511 [Verified: abstract] — IDEAS/RePEc; Working paper ERS-2013-016; Replication data ZBW
    • Important implication: OOS predictive power comes mainly from the "volatility state", that is, recognizing "we are currently in a high-volatility state", not from knowing the turning point in advance.
  • Resnick & Shoesmith (2002) "Using the Yield Curve to Time the Stock Market": extends the recession-forecasting probit to predict a bear market one month ahead, with the spread between 10-year-plus Treasuries and 3-month T-bills as the explanatory variable; the abstract says the probit strategy has a higher average excess return than an all-stock buy-and-hold, with lower total and systematic risk. FAJ 58(3):82–90, DOI 10.2469/faj.v58.n3.2540 [Verified: abstract] — CFA Institute; IDEAS
    • Details such as the 1971/01–1999/12 sample and the 50% probability threshold come from a secondary source [Secondary] — Quantpedia. Whether it is truly OOS (recursive estimation) and whether costs are deducted: could not be verified. The sample ends in 1999 and does not cover the repeated yield-curve inversions after 2000 (2006–07, 2019, 2022–24).
  • Chauvet & Potter (2000) "Coincident and leading indicators of the stock market" (J. Empirical Finance 7(1):87–111): builds bull/bear probabilities for the stock market with a dynamic factor/Markov switching model [Not verified page by page; DOI and OOS numbers could not be verified].
  • Candelon, Piplack & Straetmans (2008) "On measuring synchronization of bulls and bears: The case of East Asia" (JBF 32(6)) is a paper measuring the synchronization of bull and bear markets across markets, not a turning-point prediction paper; the original was not obtained in this search [Not verified page by page].
  • Hauptmann et al. (market regime prediction, regime-switching + logit): the bibliographic record and OOS numbers could not be obtained this time; insufficient evidence / could not be verified.

Bubble detection

  • Phillips, Shi & Yu (2015) "Testing for Multiple Bubbles: Historical Episodes of Exuberance and Collapse in the S&P 500": proposes the flexible-window recursive GSADF test and a date-stamping rule for the start and end dates of "multiple bubbles"; applied to the S&P 500 over 1871–2010, it identifies the known episodes of exuberance and collapse, whereas the PWY and CUSUM methods find far fewer episodes. IER 56(4):1043–1078, DOI 10.1111/iere.12132 [Verified: abstract] — SSRN; SMU INK
    • The method can be computed in real time (recursive, using only data known at the time), but what it detects is "explosive growth of prices relative to dividends", whose start is usually several years before the top; the specific start and end dates of each bubble in the paper could not be verified this time.
  • Phillips, Wu & Yu (2011) "Explosive Behavior in the 1990s Nasdaq: When Did Exuberance Escalate Asset Values?" (IER 52(1):201–226, DOI 10.1111/j.1468-2354.2010.00625.x) is a bubble-detection paper that targets the Nasdaq directly; as remembered, the detected explosive phase began around mid-1995 and ended in 2000–2001 [Not verified page by page; dates must be confirmed against the original] — DOI
  • Greenwood, Shleifer & You (2019) "Bubbles for Fama", JFE 131(1):20–43, DOI 10.1016/j.jfineco.2018.09.002. The numbers below are taken from the original text of NBER WP 23191 [Verified: original] — NBER w23191 PDF; Shleifer page
    • Sample: US industries 1926–2014, international industries 1985–2014 (abstract, WP p. 1).
    • After an industry rises 100% within two years, the average subsequent return is about zero, i.e. "a run-up does not predict low returns" (abstract; WP p. 13).
    • A crash is defined as a 40% drawdown within two years. As an industry's return "net of the market" rises from 50% to 100%, the crash probability rises from 20% to 53%; at 150% it rises to 80% (WP introduction pp. 4–5; p. 14, corresponding to Table 3). In the international sample it rises from 36% to 50% (WP p. 14).
    • Unconditional crash probability: 14% for any two-year period for US industries (11% after 1970), 24% internationally (WP pp. 13–14).
    • The international sample as an out-of-sample test: of 107 run-ups, 53 crashed and 54 did not (WP pp. 7–8, Table 2).
    • Trading rules (Table 8, WP pp. 26–28): selling based only on a 100% run-up, in the US sample 53% of 40 cases crashed and 47% were false alarms; at the 1-year horizon it lags the passive strategy by about 5%, because even when correctly identified, the industry on average keeps rising for about another 6 months. At the 2-year horizon, switching to the risk-free asset wins by about 10%, but "statistical significance is quite limited", and "only 40 observations". Adding a volatility condition: 20 candidates, 60% crashed; excess returns from switching to the risk-free asset at 1/2/4 years are 0%/8%/12%, with only the 4-year horizon significant.
    • In fn. 14 (WP p. 27) the authors admit that setting the threshold at the sample average of the crash group "introduces some look-ahead bias into our portfolio results".
    • The objects of study are industry portfolios (including Computer Software and Computer Hardware in the late 1990s), not the broad market index; WP pp. 9–10 list the software and hardware run-ups of the late 1990s.
  • Criticism of LPPL (Sornette): Brée & Joseph examined 11 crashes in the Hang Seng Index over 1970–2008 and found that only 7 had parameters within the claimed ranges, and they argue its mechanism argument does not hold. Brée, Challet & Peirano (Quantitative Finance 2013) point out that LPPL is hard to fit, the crash time t_c is the hardest parameter to pin down, and current probabilistic time-window estimates may be overly optimistic. A 2025 IRFA paper calls for re-examining LPPLS evidence that may be false positives caused by sample size. [Verified: abstract] — arXiv 1002.1010; Folia/Unifr; UWasa
  • Goetzmann (2016) "Bubble Investing: Learning from History" (NBER WP 21693): the conclusion as remembered is that crashes follow run-ups only slightly more often than unconditionally, and crashes often happen without a run-up [Not verified page by page] — NBER w21693. GSY 2019 itself (around WP p. 13) also notes that Goetzmann finds a higher crash probability, but only modestly higher.

Inferences

  • "Can tops or bottoms be called in real time?": the answer is no. All verified evidence falls into three categories: (1) ex-post dating; (2) predicting the probability of being in a bear state in the next period (mainly from the term spread, inflation and the volatility state), which can push the bear probability above its unconditional value, but not a single verified paper reports stable, low-false-alarm, after-cost prediction of turning-point dates; (3) crash probability rises after run-ups, but the timing is unpredictable (GSY: after correct identification, prices on average keep rising for 6 months).
  • How these methods differ from "selling at relative highs and buying at relative lows": in substance they are "risk-state switching" (reducing exposure when volatility is high or the yield curve is inverted), and the gains come mainly from avoiding the deep-decline phase of bear markets, not from selling at the top. Kole & van Dijk's conclusion (variance is the key to forecasting) supports this reading.
  • Out-of-sample evidence is concentrated on the S&P 500 or the broad US market, at monthly frequency, from the 1950s to the 2000s. Resnick-Shoesmith's sample stops in 1999; Chen's stops in 2007. The yield curve's prolonged inversion in 2022–2024 while stocks rallied strongly was a major false alarm for spread rules, so their applicability is questionable (this is an inference; the main report should verify it independently with FRED T10Y3M).
  • For QQQ: GSY's industry run-up rule is conceptually the closest fit to a tech-concentrated index, but it works at the industry level, has only 40 events, its threshold has look-ahead bias, and the average return after a run-up is about zero. Its implication is "tail risk rises after a run-up", not "you should sell after a run-up".

Gaps

  • OOS QPS, AUC, hit rate, false-alarm rate, and whether trading excess returns are after costs for Nyberg 2013, Chen 2009 and Resnick-Shoesmith 2002: full texts are paywalled; could not be verified this time.
  • OOS numbers for Maheu-McCurdy-Song 2012, Chauvet-Potter 2000, Hauptmann et al.: not obtained.
  • No peer-reviewed paper was found that systematically compares multiple bear-market predictors by ROC/AUC and covers 2008–2024 (one may exist, but this search did not find it).
  • The start and end dates of the dot-com-era S&P 500 bubble in PSY 2015 and the Nasdaq dates in PWY 2011: could not be verified.

Q2 Evidence specific to the Nasdaq-100/QQQ and the risk of extrapolating from the S&P 500

Takeaway

This search found no peer-reviewed study that times the Nasdaq-100 or Nasdaq Composite index out of sample (trend, valuation, VXN) and beats buy-and-hold. The top-journal Nasdaq literature focuses on after-the-fact explanations of the late-1990s bubble: Pastor-Veronesi argue valuations can be rationalized; Ofek-Richardson explain it through short-sale constraints and lockups; Brunnermeier-Nagel find hedge funds "rode the bubble" rather than resisting it; PWY 2011 does bubble detection. Extrapolating from the S&P 500 to QQQ carries real risk: sector concentration, intangible capital distorting earnings and CAPE, short samples (the Nasdaq-100 since 1985, QQQ since 1999), and very few events (in practice only two big bear markets, 2000–02 and 2022).

Cited Findings

  • Pastor & Veronesi (2006) "Was There a Nasdaq Bubble in the Late 1990s?": a firm's fundamental value rises with "uncertainty about average future profitability", and that uncertainty was unusually high in the late 1990s; after calibrating the model, the required uncertainty can match both the high price level and the high volatility of the Nasdaq peak, so the peak prices can be rationalized within their rational model; the effect of uncertainty is greatest when the equity premium is low. The authors do not claim there was no bubble, but argue it is not obvious. JFE 81(1):61–100, DOI 10.1016/j.jfineco.2005.05.009 [Verified: abstract] — NBER w10581; Advisor.ca commentary
  • Brunnermeier & Nagel (2004) "Hedge Funds and the Technology Bubble": hedge funds were not a price-correcting force but instead were heavily invested in technology stocks; they captured the upturn and cut their positions in individual stocks before those stocks fell, thereby avoiding most of the downturn; the results are consistent with models in which "rational investors choose to ride the bubble because of predictable sentiment and limits to arbitrage". JF 59(5):2013–2040, DOI 10.1111/j.1540-6261.2004.00690.x [Verified: abstract] — Princeton page; Stanford GSB
    • Note: the hedge funds' timing was at the individual-stock level and relied on information and flow advantages; it does not show that retail investors can time QQQ with public indicators.
  • Ofek & Richardson (2003) "DotCom Mania: The Rise and Fall of Internet Stock Prices": explains the high prices and collapse of internet stocks through short-sale constraints, concentrated retail ownership, and the increase in supply when lockups expired. JF 58(3):1113–1137, DOI 10.1111/1540-6261.00560 [Not verified page by page; specific P/E ratios and sample size not verified] — DOI
  • Phillips, Wu & Yu (2011) does real-time bubble detection on the Nasdaq (see Q1) [Not verified page by page] — DOI
  • The long-run evidence for moving-average/trend rules is all on the S&P Composite or the DJIA: Zakamulin, using monthly S&P data for 1860–2009, finds that MA timing has lower risk and usually a higher Sharpe ratio but a lower average return; later research finds no statistically significant outperformance in the second half of the sample, and many false signals in both bull and bear trends. This search found no peer-reviewed OOS study using the Nasdaq-100/Composite as its sample. [Secondary; some versions of Zakamulin are working papers] — SSRN 2677212; CXO Advisory summary; UiA thesis (rules profitable in sample perform poorly out of sample; results depend heavily on the split point)
  • Intangible capital distorts accounting values: Lev & Srivastava argue value investing has broadly not made money for nearly 30 years (with a brief exception after the dot-com crash), partly because accounting expenses investment in R&D, IT, brands and human capital, so value is systematically misidentified; an adjusted value strategy that capitalizes R&D and SG&A has performed noticeably better since the late 1980s, though this does not fully explain it either. Critical Finance Review version (2021/2022) [Verified: abstract + Secondary] — CFR PDF; Institutional Investor
  • Peters & Taylor (2017) "Intangible capital and the investment-q relation": once intangible capital is included in Tobin's q, q's power to explain investment improves markedly. JFE 123(2):251–272, DOI 10.1016/j.jfineco.2016.03.011 [Not verified page by page] — DOI
  • Eisfeldt, Kim & Papanikolaou "Intangible Value" (NBER WP 28056; Critical Finance Review 2022): an improved HML that includes intangible capital in book value performs better [Not verified page by page] — NBER w28056
  • Dichev (2007): NASDAQ dollar-weighted returns over 1973–2002 were 5.3 percentage points per year lower than buy-and-hold (GM 9.6% vs DW 4.3%), NYSE/AMEX over 1926–2002 was 1.3 percentage points lower, and 19 international markets were 1.5 percentage points lower on average, indicating that Nasdaq investors as a whole had particularly poor "entry and exit timing". AER 97(1):386–401, DOI 10.1257/aer.97.1.386 [Numbers from abstract + Secondary] — City Univ. follow-up paper tables; U-M press release
    • Rebuttal: Keswani & Stolin (2008, Finance Research Letters 5(4):228–235) argue that the NYSE/AMEX gap is actually negative in the last two-thirds of the period, that the international results are affected by the expansion of Datastream coverage, and that Dichev's conclusion is not robust (though they do not directly address the NASDAQ figure) — IDEAS
  • FRED's NASDAQ100 series is provided by Nasdaq, Inc., as daily closes, and is marked "Copyrighted: Pre-Approval Required"; related series include VXNCLS, NASDAQCOM and SP500 [Verified: original (FRED page)] — FRED NASDAQ100

Inferences

  • List of extrapolation risks (inferences, based on the literature above):
    1. Sector concentration: the Nasdaq-100 excludes financials, has heavy technology and communications weights, and its returns are driven by a few mega-cap stocks. The relationship between S&P 500 macro predictors (term spread, inflation) and Nasdaq-100 returns may differ; GSY show that crash probability rises after industry-level run-ups, a mechanism that applies more directly to the sector-concentrated QQQ than to VOO.
    2. Distorted valuation indicators: expensing intangible investment depresses earnings and book value, so the Nasdaq-100's P/E, CAPE and P/B stay high for long periods (Lev-Srivastava; Peters-Taylor; EKP). Applying the historical S&P 500 CAPE thresholds to QQQ is very likely to produce a long-running false "overvalued, reduce exposure" signal.
    3. Very short samples: the Nasdaq-100 since 1985, QQQ since 1999, with only two big bear markets, 2000–02 (roughly more than -80%) and 2022 (roughly more than -30%), plus 2008. Any turning-point rule backtested on QQQ has extremely few degrees of freedom and a very high risk of overfitting (the specific drawdown figures were not verified this time).
    4. Index rules and survivorship bias: the Nasdaq-100 selects stocks by market cap and reconstitutes periodically, so the index itself has a built-in momentum characteristic; a "synthetic history" backfilled from today's constituents has survivorship bias (an inference; no dedicated study was found).
    5. Mega-cap concentration: the top ten weights are high, index returns are closely tied to a few companies' earnings reports, and the explanatory power of macro timing signals may be diluted by single-stock events.
  • Pastor-Veronesi's implication for timing: if even the 2000 Nasdaq peak can be rationalized in a rational model, "calling the top on valuation" lacks a criterion ex ante.
  • Dichev's Nasdaq gap of 5.3%/year is the most direct academic evidence of "the actual cost of ordinary investors timing the Nasdaq" (though subject to Keswani-Stolin's challenge to the method's robustness).

Gaps

  • No peer-reviewed OOS timing study using the Nasdaq-100/Composite/QQQ as its sample was found (none for trend, valuation or VXN). Such studies may exist in lower-tier journals or working papers, but this search did not obtain them, so this should be treated as "insufficient evidence".
  • Academic evidence on VXN predicting Nasdaq-100 returns or bear markets: could not be verified.
  • Historical index-level earnings and P/E series for the Nasdaq-100: no free official source going back to 1985 was found (see Q4).
  • Quantitative research on survivorship bias in the Nasdaq-100 index rules: not found.

Q3 Retail implementation realities: trading costs, taxes, and investors' actual timing record

Takeaway

Academic evidence consistently shows that ordinary investors' actual entry and exit timing eats into returns (Dichev: 5.3 percentage points per year for the Nasdaq; Friesen-Sapp: investor timing costs fund investors about 1.56% per year, to be verified). For US taxable accounts, timing also incurs capital gains tax; for Taiwan residents, capital gains on US-listed ETFs count as overseas income, which is included in basic income only when the household's overseas income reaches NT$1 million, and the alternative minimum tax only starts once basic income exceeds NT$7.5 million. So for most Taiwanese retail investors the "tax drag from timing" is far smaller than for US investors, but this does not change the evidence on whether timing can generate excess returns.

Cited Findings

  • Dichev (2007): see Q2; the gap between dollar-weighted and buy-and-hold returns reflects the timing of investors' inflows and outflows (buying high, selling low) — U-M press release; DOI
  • Friesen & Sapp (2007) "Mutual fund flows and investor returns: An empirical examination of fund investor timing ability", JBF 31(9):2796–2816, DOI 10.1016/j.jbankfin.2007.01.024. The conclusion as remembered: fund investors' timing decisions reduce returns by about 1.56% per year, concentrated mainly in funds with strong recent performance and large inflows [Not verified page by page; numbers must be checked against the original] — DOI
  • Dalbar QAIB-type reports are not academic research, their methodology has been repeatedly criticized, and they serve only as leads (no figures are cited this time).
  • The GSY 2019 trading rules do not account for trading costs or taxes, and the threshold has look-ahead bias (WP p. 27, fn. 14) [Verified: original] — NBER w23191
  • Taiwan's alternative minimum tax (individuals): if the same filing household's total overseas income for the year reaches NT$1 million, all of it must be included in basic income; the basic income exemption is NT$7.5 million (the same for tax years 2024 and 2025); the basic tax is (basic income − NT$7.5 million) × 20%, compared with the regular income tax with the higher of the two applying; income from overseas stocks is recognized in the year of the settlement date and must be reported once realized, whether or not it is remitted back [Secondary (news + personal finance sites, citing the National Taxation Bureau's explanation); check the Ministry of Finance's original text before filing] — United Daily News (udn.com); United Daily News (udn.com) 2; Money101
  • The US withholds 30% on dividends paid to nonresident aliens (no US-Taiwan tax treaty): this is general tax knowledge; the original regulatory page was not found this time, [could not be verified]; the latest status of the US-Taiwan double-taxation-relief legislation could not be verified either.

Inferences

  • On taxes: the 30% withholding on US ETF dividends is roughly the same under timing and buy-and-hold (both receive dividends; if cash or a bond ETF is held during the timing period, the dividend withholding is replaced by the treatment of interest-related income, and the difference is small), so it does not affect the judgment of "whether timing works".
  • Capital gains for Taiwan residents: if the household's overseas income is below NT$1 million, or basic income is below NT$7.5 million, frequently realizing gains effectively produces no extra tax; but once a large investor exceeds the thresholds, the "early realization of gains" caused by timing turns deferral into a 20% alternative minimum tax in that year. This differs from US investors, for whom "long-term holding allows deferral and the long-term capital gains rate".
  • Trading costs: QQQ and VOO have very small bid-ask spreads and most US brokers charge zero commission, so the direct cost of switching monthly is low; the real costs are: (a) currency conversion costs (TWD↔USD) and sub-brokerage fees (if going through a Taiwanese broker); (b) the opportunity cost of missing big up days while out of the market; (c) behavioral costs (the Dichev and Friesen-Sapp evidence). The specific rates in (a) vary by broker and were not verified.
  • Most academic timing backtests are before tax, some are before costs, and they are computed on indices rather than ETFs; returns achievable by retail investors should be discounted accordingly.

Gaps

  • Academic papers that directly address "tactical timing after tax vs buy-and-hold" (for example, tax-aware timing research for US taxable investors): no verifiable bibliographic record was found in this search; insufficient evidence.
  • The exact Friesen-Sapp figures were not verified.
  • Actual rates for Taiwanese sub-brokerage and for currency conversion and commissions at overseas brokers were not verified; the latest status of a US-Taiwan tax agreement or related legislation under the Taiwan Relations Act was not verified.

Q4 Free data sources and publication lags

Takeaway

Free data on the S&P 500 side is complete (Shiller monthly data back to 1871, Goyal's predictor variables, FRED spreads and VIX); on the Nasdaq-100 side only prices and VXN are easy to obtain, and index-level earnings/valuation history lacks a free, long-run official source, which in itself limits how testable valuation-based timing of QQQ is.

Cited Findings

  • FRED: NASDAQ100 (daily closes, source Nasdaq, Inc., marked Copyrighted: Pre-Approval Required); related series include VXNCLS, NASDAQCOM and SP500 [Verified: original] — FRED NASDAQ100
    • The SP500 series on FRED provides only the last 10 years of daily data because of licensing (from memory, not verified); the start date of NASDAQ100 was not shown on the page this time (as remembered, around 1986), and could not be verified.
  • Phillips-Shi-Yu use S&P 500 price/dividend data for 1871–2010 (i.e., Shiller-type data) — SSRN
  • The Goyal-Welch-Zafirov 2024 data extend through the end of 2021; Goyal's personal website provides the predictor-variable data (URL not verified this time) — SFI
  • The Kole & van Dijk 2017 replication data are public in the ZBW Journal Data Archive — ZBW

Inferences

  • Sources the main report should list (the URLs are common knowledge and were not opened one by one this time): Shiller's "Irrational Exuberance" data page (S&P monthly prices, dividends, earnings, CAPE); FRED (T10Y3M, T10Y2Y, VIXCLS, VXNCLS, SP500, NASDAQCOM, NASDAQ100, CPI, Z.1 Financial Accounts); Amit Goyal's website (GW predictor variables); Jeffrey Wurgler's website (Baker-Wurgler sentiment index, updated with a lag); CBOE (VIX, VXN history); Yahoo Finance/Stooq (dividend-adjusted ETF prices).
  • Publication lags are critical for OOS testing: CPI lags about half a month, Shiller earnings lag several months, Z.1 lags about a quarter, and the Baker-Wurgler sentiment index lags even longer. If a backtest uses the "current month" values, it introduces look-ahead bias (an inference; the specific lag periods were not verified one by one).
  • Valuation-based timing of the Nasdaq-100 requires aggregating constituent earnings yourself or using paid data (Bloomberg, FactSet, recent WisdomTree/Invesco data), and both historical depth and consistency are problematic, so testability is low.

Gaps

  • A free Nasdaq-100 index-level EPS/P/E series going back to 1985: not found.
  • The exact start dates of FRED NASDAQ100 and VXNCLS: not shown on the page this time; could not be verified.
  • The precise publication lag in days for each data source: not verified one by one.

Further reading

The other four articles in this series:

Related pages on this site:

Frequently asked questions

Is there a method that can catch stock-market tops or bottoms in real time?

The note's answer is no. The verified evidence falls into only three categories: ex-post dating of bull/bear cycles; predicting the probability of being in a bear state next period (mainly from the term spread, inflation and the volatility state), though not a single verified paper reports stable, low-false-alarm, after-cost prediction of turning-point dates; and the finding that crash probability rises after run-ups, while the timing is unpredictable. Bubble detection (GSADF) detects that an explosive run-up is already under way; it does not predict the top.

Can an inverted yield curve predict US bear markets?

Chen (2009) finds that the yield-curve spread and inflation are the most useful bear-market predictors, both in sample and out of sample, but the specific OOS numbers could not be verified. Resnick & Shoesmith (2002)'s sample stops in 1999 and does not cover the repeated inversions after 2000; the note further infers that the yield curve's prolonged inversion in 2022–2024 while stocks rallied strongly was a major false alarm for spread rules.

Does a crash always follow an industry run-up?

Greenwood, Shleifer & You (2019) find that after an industry rises 100% within two years, the average subsequent return is about zero; as an industry's return net of the market rises from 50% to 100%, the crash probability rises from 20% to 53%, and to 80% at 150%. But even when correctly identified, the industry on average keeps rising for about another 6 months; the trading rule has only 40 observations, its significance is limited, and the authors admit the threshold has look-ahead bias.

How much does it actually cost ordinary investors to time the Nasdaq themselves?

Dichev (2007) finds that NASDAQ dollar-weighted returns over 1973–2002 were 5.3 percentage points per year lower than buy-and-hold (9.6% vs 4.3%), indicating that Nasdaq investors as a whole had particularly poor entry and exit timing; but Keswani & Stolin (2008) question the robustness of Dichev's conclusion (without directly addressing the NASDAQ figure).

Do Taiwanese investors who trade US ETFs frequently owe tax on capital gains?

According to the secondary sources cited in the note: overseas income is included in basic income only when the same filing household's total overseas income for the year reaches NT$1 million, the alternative minimum tax only starts once basic income exceeds NT$7.5 million, and the basic tax is (basic income − NT$7.5 million) × 20%. So the tax drag from timing for most Taiwanese retail investors is far smaller than for US investors, but this does not change the evidence on whether timing can generate excess returns; check the Ministry of Finance's original text before filing.

Why can't QQQ timing backtest results be trusted at face value?

The extrapolation risks listed in the note include: the Nasdaq-100 since 1985 and QQQ since 1999, with only two big bear markets, 2000–02 and 2022, plus 2008, so degrees of freedom are extremely low; expensing intangible investment keeps P/E and CAPE high for long periods; sector and mega-cap concentration; and survivorship bias in synthetic histories backfilled from today's constituents. This search found no peer-reviewed study that times the Nasdaq-100 out of sample and beats buy-and-hold.

More articles

← All articles Open today's market dashboard →

Contents